
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (floor w) dY.u))
(t_2 (* (floor w) dX.u))
(t_3 (+ (* t_2 t_2) (* t_0 t_0)))
(t_4 (* (floor h) dY.v))
(t_5 (+ (* t_1 t_1) (* t_4 t_4)))
(t_6 (/ 1.0 (sqrt (fmax t_3 t_5)))))
(if (>= t_3 t_5) (* t_6 t_2) (* t_6 t_1))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = floorf(w) * dY_46_u;
float t_2 = floorf(w) * dX_46_u;
float t_3 = (t_2 * t_2) + (t_0 * t_0);
float t_4 = floorf(h) * dY_46_v;
float t_5 = (t_1 * t_1) + (t_4 * t_4);
float t_6 = 1.0f / sqrtf(fmaxf(t_3, t_5));
float tmp;
if (t_3 >= t_5) {
tmp = t_6 * t_2;
} else {
tmp = t_6 * t_1;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(floor(w) * dX_46_u) t_3 = Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) t_4 = Float32(floor(h) * dY_46_v) t_5 = Float32(Float32(t_1 * t_1) + Float32(t_4 * t_4)) t_6 = Float32(Float32(1.0) / sqrt(((t_3 != t_3) ? t_5 : ((t_5 != t_5) ? t_3 : max(t_3, t_5))))) tmp = Float32(0.0) if (t_3 >= t_5) tmp = Float32(t_6 * t_2); else tmp = Float32(t_6 * t_1); end return tmp end
function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(h) * dX_46_v; t_1 = floor(w) * dY_46_u; t_2 = floor(w) * dX_46_u; t_3 = (t_2 * t_2) + (t_0 * t_0); t_4 = floor(h) * dY_46_v; t_5 = (t_1 * t_1) + (t_4 * t_4); t_6 = single(1.0) / sqrt(max(t_3, t_5)); tmp = single(0.0); if (t_3 >= t_5) tmp = t_6 * t_2; else tmp = t_6 * t_1; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_3 := t\_2 \cdot t\_2 + t\_0 \cdot t\_0\\
t_4 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_5 := t\_1 \cdot t\_1 + t\_4 \cdot t\_4\\
t_6 := \frac{1}{\sqrt{\mathsf{max}\left(t\_3, t\_5\right)}}\\
\mathbf{if}\;t\_3 \geq t\_5:\\
\;\;\;\;t\_6 \cdot t\_2\\
\mathbf{else}:\\
\;\;\;\;t\_6 \cdot t\_1\\
\end{array}
\end{array}
Sampling outcomes in binary32 precision:
Herbie found 6 alternatives:
| Alternative | Accuracy | Speedup |
|---|
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (floor w) dY.u))
(t_2 (* (floor w) dX.u))
(t_3 (+ (* t_2 t_2) (* t_0 t_0)))
(t_4 (* (floor h) dY.v))
(t_5 (+ (* t_1 t_1) (* t_4 t_4)))
(t_6 (/ 1.0 (sqrt (fmax t_3 t_5)))))
(if (>= t_3 t_5) (* t_6 t_2) (* t_6 t_1))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = floorf(w) * dY_46_u;
float t_2 = floorf(w) * dX_46_u;
float t_3 = (t_2 * t_2) + (t_0 * t_0);
float t_4 = floorf(h) * dY_46_v;
float t_5 = (t_1 * t_1) + (t_4 * t_4);
float t_6 = 1.0f / sqrtf(fmaxf(t_3, t_5));
float tmp;
if (t_3 >= t_5) {
tmp = t_6 * t_2;
} else {
tmp = t_6 * t_1;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(floor(w) * dX_46_u) t_3 = Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) t_4 = Float32(floor(h) * dY_46_v) t_5 = Float32(Float32(t_1 * t_1) + Float32(t_4 * t_4)) t_6 = Float32(Float32(1.0) / sqrt(((t_3 != t_3) ? t_5 : ((t_5 != t_5) ? t_3 : max(t_3, t_5))))) tmp = Float32(0.0) if (t_3 >= t_5) tmp = Float32(t_6 * t_2); else tmp = Float32(t_6 * t_1); end return tmp end
function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(h) * dX_46_v; t_1 = floor(w) * dY_46_u; t_2 = floor(w) * dX_46_u; t_3 = (t_2 * t_2) + (t_0 * t_0); t_4 = floor(h) * dY_46_v; t_5 = (t_1 * t_1) + (t_4 * t_4); t_6 = single(1.0) / sqrt(max(t_3, t_5)); tmp = single(0.0); if (t_3 >= t_5) tmp = t_6 * t_2; else tmp = t_6 * t_1; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_3 := t\_2 \cdot t\_2 + t\_0 \cdot t\_0\\
t_4 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_5 := t\_1 \cdot t\_1 + t\_4 \cdot t\_4\\
t_6 := \frac{1}{\sqrt{\mathsf{max}\left(t\_3, t\_5\right)}}\\
\mathbf{if}\;t\_3 \geq t\_5:\\
\;\;\;\;t\_6 \cdot t\_2\\
\mathbf{else}:\\
\;\;\;\;t\_6 \cdot t\_1\\
\end{array}
\end{array}
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(t_1 (* dY.u (floor w)))
(t_2 (+ (pow t_1 2.0) (pow (* dY.v (floor h)) 2.0)))
(t_3 (+ (pow (* (floor h) dX.v) 2.0) (pow t_0 2.0)))
(t_4 (sqrt (fmax t_3 t_2))))
(if (>= t_3 t_2) (/ t_0 t_4) (/ t_1 t_4))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(w) * dX_46_u;
float t_1 = dY_46_u * floorf(w);
float t_2 = powf(t_1, 2.0f) + powf((dY_46_v * floorf(h)), 2.0f);
float t_3 = powf((floorf(h) * dX_46_v), 2.0f) + powf(t_0, 2.0f);
float t_4 = sqrtf(fmaxf(t_3, t_2));
float tmp;
if (t_3 >= t_2) {
tmp = t_0 / t_4;
} else {
tmp = t_1 / t_4;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(w) * dX_46_u) t_1 = Float32(dY_46_u * floor(w)) t_2 = Float32((t_1 ^ Float32(2.0)) + (Float32(dY_46_v * floor(h)) ^ Float32(2.0))) t_3 = Float32((Float32(floor(h) * dX_46_v) ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) t_4 = sqrt(((t_3 != t_3) ? t_2 : ((t_2 != t_2) ? t_3 : max(t_3, t_2)))) tmp = Float32(0.0) if (t_3 >= t_2) tmp = Float32(t_0 / t_4); else tmp = Float32(t_1 / t_4); end return tmp end
function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(w) * dX_46_u; t_1 = dY_46_u * floor(w); t_2 = (t_1 ^ single(2.0)) + ((dY_46_v * floor(h)) ^ single(2.0)); t_3 = ((floor(h) * dX_46_v) ^ single(2.0)) + (t_0 ^ single(2.0)); t_4 = sqrt(max(t_3, t_2)); tmp = single(0.0); if (t_3 >= t_2) tmp = t_0 / t_4; else tmp = t_1 / t_4; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_2 := {t\_1}^{2} + {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2}\\
t_3 := {\left(\left\lfloor h\right\rfloor \cdot dX.v\right)}^{2} + {t\_0}^{2}\\
t_4 := \sqrt{\mathsf{max}\left(t\_3, t\_2\right)}\\
\mathbf{if}\;t\_3 \geq t\_2:\\
\;\;\;\;\frac{t\_0}{t\_4}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_1}{t\_4}\\
\end{array}
\end{array}
Initial program 77.1%
lift-*.f32N/A
lift-/.f32N/A
associate-*l/N/A
frac-2negN/A
Applied rewrites77.3%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3277.3
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3277.3
Applied rewrites77.3%
lift-+.f32N/A
+-commutativeN/A
lift-*.f32N/A
pow2N/A
lower-pow.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f32N/A
lift-+.f3277.3
Applied rewrites77.3%
lift-*.f32N/A
*-commutativeN/A
lift-/.f32N/A
frac-2negN/A
metadata-evalN/A
associate-*r/N/A
lower-/.f32N/A
Applied rewrites77.4%
Final simplification77.4%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(t_1 (* dY.u (floor w)))
(t_2 (+ (pow t_1 2.0) (pow (* dY.v (floor h)) 2.0)))
(t_3 (+ (pow (* (floor h) dX.v) 2.0) (pow t_0 2.0)))
(t_4 (sqrt (fmax t_3 t_2))))
(if (>= t_3 t_2) (* t_0 (/ 1.0 t_4)) (/ t_1 t_4))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(w) * dX_46_u;
float t_1 = dY_46_u * floorf(w);
float t_2 = powf(t_1, 2.0f) + powf((dY_46_v * floorf(h)), 2.0f);
float t_3 = powf((floorf(h) * dX_46_v), 2.0f) + powf(t_0, 2.0f);
float t_4 = sqrtf(fmaxf(t_3, t_2));
float tmp;
if (t_3 >= t_2) {
tmp = t_0 * (1.0f / t_4);
} else {
tmp = t_1 / t_4;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(w) * dX_46_u) t_1 = Float32(dY_46_u * floor(w)) t_2 = Float32((t_1 ^ Float32(2.0)) + (Float32(dY_46_v * floor(h)) ^ Float32(2.0))) t_3 = Float32((Float32(floor(h) * dX_46_v) ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) t_4 = sqrt(((t_3 != t_3) ? t_2 : ((t_2 != t_2) ? t_3 : max(t_3, t_2)))) tmp = Float32(0.0) if (t_3 >= t_2) tmp = Float32(t_0 * Float32(Float32(1.0) / t_4)); else tmp = Float32(t_1 / t_4); end return tmp end
function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(w) * dX_46_u; t_1 = dY_46_u * floor(w); t_2 = (t_1 ^ single(2.0)) + ((dY_46_v * floor(h)) ^ single(2.0)); t_3 = ((floor(h) * dX_46_v) ^ single(2.0)) + (t_0 ^ single(2.0)); t_4 = sqrt(max(t_3, t_2)); tmp = single(0.0); if (t_3 >= t_2) tmp = t_0 * (single(1.0) / t_4); else tmp = t_1 / t_4; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_2 := {t\_1}^{2} + {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2}\\
t_3 := {\left(\left\lfloor h\right\rfloor \cdot dX.v\right)}^{2} + {t\_0}^{2}\\
t_4 := \sqrt{\mathsf{max}\left(t\_3, t\_2\right)}\\
\mathbf{if}\;t\_3 \geq t\_2:\\
\;\;\;\;t\_0 \cdot \frac{1}{t\_4}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_1}{t\_4}\\
\end{array}
\end{array}
Initial program 77.1%
lift-*.f32N/A
lift-/.f32N/A
associate-*l/N/A
frac-2negN/A
Applied rewrites77.3%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3277.3
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3277.3
Applied rewrites77.3%
lift-+.f32N/A
+-commutativeN/A
lift-*.f32N/A
pow2N/A
lower-pow.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f32N/A
lift-+.f3277.3
Applied rewrites77.3%
lift-/.f32N/A
clear-numN/A
associate-/r/N/A
metadata-evalN/A
neg-mul-1N/A
neg-sub0N/A
lower--.f3277.3
lift-*.f32N/A
*-commutativeN/A
lower-*.f3277.3
Applied rewrites77.3%
Final simplification77.3%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (pow (floor h) 2.0) dX.v))
(t_1 (* dY.v (floor h)))
(t_2 (* (floor h) dX.v))
(t_3 (* (floor w) dX.u))
(t_4 (pow t_3 2.0))
(t_5 (* t_3 t_3))
(t_6 (* dY.u (floor w)))
(t_7 (+ (* t_1 t_1) (* t_6 t_6)))
(t_8 (pow t_6 2.0))
(t_9 (/ 1.0 (sqrt (fmax (+ (* t_2 t_2) t_5) t_7))))
(t_10 (* t_9 t_6)))
(if (<= dY.v 1000000000.0)
(if (>= (+ (pow t_2 2.0) t_4) t_8)
(* (/ 1.0 (sqrt (fmax (+ (* t_0 dX.v) t_5) t_7))) t_3)
t_10)
(if (>= (fma t_0 dX.v t_4) t_8) (* t_9 t_3) t_10))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = powf(floorf(h), 2.0f) * dX_46_v;
float t_1 = dY_46_v * floorf(h);
float t_2 = floorf(h) * dX_46_v;
float t_3 = floorf(w) * dX_46_u;
float t_4 = powf(t_3, 2.0f);
float t_5 = t_3 * t_3;
float t_6 = dY_46_u * floorf(w);
float t_7 = (t_1 * t_1) + (t_6 * t_6);
float t_8 = powf(t_6, 2.0f);
float t_9 = 1.0f / sqrtf(fmaxf(((t_2 * t_2) + t_5), t_7));
float t_10 = t_9 * t_6;
float tmp_1;
if (dY_46_v <= 1000000000.0f) {
float tmp_2;
if ((powf(t_2, 2.0f) + t_4) >= t_8) {
tmp_2 = (1.0f / sqrtf(fmaxf(((t_0 * dX_46_v) + t_5), t_7))) * t_3;
} else {
tmp_2 = t_10;
}
tmp_1 = tmp_2;
} else if (fmaf(t_0, dX_46_v, t_4) >= t_8) {
tmp_1 = t_9 * t_3;
} else {
tmp_1 = t_10;
}
return tmp_1;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32((floor(h) ^ Float32(2.0)) * dX_46_v) t_1 = Float32(dY_46_v * floor(h)) t_2 = Float32(floor(h) * dX_46_v) t_3 = Float32(floor(w) * dX_46_u) t_4 = t_3 ^ Float32(2.0) t_5 = Float32(t_3 * t_3) t_6 = Float32(dY_46_u * floor(w)) t_7 = Float32(Float32(t_1 * t_1) + Float32(t_6 * t_6)) t_8 = t_6 ^ Float32(2.0) t_9 = Float32(Float32(1.0) / sqrt(((Float32(Float32(t_2 * t_2) + t_5) != Float32(Float32(t_2 * t_2) + t_5)) ? t_7 : ((t_7 != t_7) ? Float32(Float32(t_2 * t_2) + t_5) : max(Float32(Float32(t_2 * t_2) + t_5), t_7))))) t_10 = Float32(t_9 * t_6) tmp_1 = Float32(0.0) if (dY_46_v <= Float32(1000000000.0)) tmp_2 = Float32(0.0) if (Float32((t_2 ^ Float32(2.0)) + t_4) >= t_8) tmp_2 = Float32(Float32(Float32(1.0) / sqrt(((Float32(Float32(t_0 * dX_46_v) + t_5) != Float32(Float32(t_0 * dX_46_v) + t_5)) ? t_7 : ((t_7 != t_7) ? Float32(Float32(t_0 * dX_46_v) + t_5) : max(Float32(Float32(t_0 * dX_46_v) + t_5), t_7))))) * t_3); else tmp_2 = t_10; end tmp_1 = tmp_2; elseif (fma(t_0, dX_46_v, t_4) >= t_8) tmp_1 = Float32(t_9 * t_3); else tmp_1 = t_10; end return tmp_1 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := {\left(\left\lfloor h\right\rfloor \right)}^{2} \cdot dX.v\\
t_1 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_2 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_4 := {t\_3}^{2}\\
t_5 := t\_3 \cdot t\_3\\
t_6 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_7 := t\_1 \cdot t\_1 + t\_6 \cdot t\_6\\
t_8 := {t\_6}^{2}\\
t_9 := \frac{1}{\sqrt{\mathsf{max}\left(t\_2 \cdot t\_2 + t\_5, t\_7\right)}}\\
t_10 := t\_9 \cdot t\_6\\
\mathbf{if}\;dY.v \leq 1000000000:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;{t\_2}^{2} + t\_4 \geq t\_8:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_0 \cdot dX.v + t\_5, t\_7\right)}} \cdot t\_3\\
\mathbf{else}:\\
\;\;\;\;t\_10\\
\end{array}\\
\mathbf{elif}\;\mathsf{fma}\left(t\_0, dX.v, t\_4\right) \geq t\_8:\\
\;\;\;\;t\_9 \cdot t\_3\\
\mathbf{else}:\\
\;\;\;\;t\_10\\
\end{array}
\end{array}
if dY.v < 1e9Initial program 78.4%
Taylor expanded in dY.u around inf
*-commutativeN/A
unpow2N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3269.5
Applied rewrites69.5%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3269.5
Applied rewrites69.5%
lift-+.f32N/A
+-commutativeN/A
lift-*.f32N/A
pow2N/A
lower-pow.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-+.f3269.5
Applied rewrites69.5%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
swap-sqrN/A
unpow2N/A
lift-pow.f32N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f3269.6
Applied rewrites69.6%
if 1e9 < dY.v Initial program 70.6%
Taylor expanded in dY.u around inf
*-commutativeN/A
unpow2N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3239.7
Applied rewrites39.7%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3239.7
Applied rewrites39.7%
lift-+.f32N/A
+-commutativeN/A
lift-*.f32N/A
pow2N/A
lower-pow.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-+.f3239.7
Applied rewrites39.7%
Taylor expanded in w around 0
+-commutativeN/A
*-commutativeN/A
unpow2N/A
associate-*r*N/A
lower-fma.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f32N/A
unpow2N/A
unpow2N/A
unswap-sqrN/A
unpow2N/A
lower-pow.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-floor.f3252.5
Applied rewrites52.5%
Final simplification65.5%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.v (floor h)))
(t_1 (* dY.u (floor w)))
(t_2 (+ (* t_0 t_0) (* t_1 t_1)))
(t_3 (* (floor h) dX.v))
(t_4 (* (floor w) dX.u))
(t_5 (* t_4 t_4)))
(if (>= (+ (pow t_3 2.0) (pow t_4 2.0)) (pow t_1 2.0))
(*
(/ 1.0 (sqrt (fmax (+ (* (* (pow (floor h) 2.0) dX.v) dX.v) t_5) t_2)))
t_4)
(* (/ 1.0 (sqrt (fmax (+ (* t_3 t_3) t_5) t_2))) t_1))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = dY_46_v * floorf(h);
float t_1 = dY_46_u * floorf(w);
float t_2 = (t_0 * t_0) + (t_1 * t_1);
float t_3 = floorf(h) * dX_46_v;
float t_4 = floorf(w) * dX_46_u;
float t_5 = t_4 * t_4;
float tmp;
if ((powf(t_3, 2.0f) + powf(t_4, 2.0f)) >= powf(t_1, 2.0f)) {
tmp = (1.0f / sqrtf(fmaxf((((powf(floorf(h), 2.0f) * dX_46_v) * dX_46_v) + t_5), t_2))) * t_4;
} else {
tmp = (1.0f / sqrtf(fmaxf(((t_3 * t_3) + t_5), t_2))) * t_1;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(dY_46_v * floor(h)) t_1 = Float32(dY_46_u * floor(w)) t_2 = Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) t_3 = Float32(floor(h) * dX_46_v) t_4 = Float32(floor(w) * dX_46_u) t_5 = Float32(t_4 * t_4) tmp = Float32(0.0) if (Float32((t_3 ^ Float32(2.0)) + (t_4 ^ Float32(2.0))) >= (t_1 ^ Float32(2.0))) tmp = Float32(Float32(Float32(1.0) / sqrt(((Float32(Float32(Float32((floor(h) ^ Float32(2.0)) * dX_46_v) * dX_46_v) + t_5) != Float32(Float32(Float32((floor(h) ^ Float32(2.0)) * dX_46_v) * dX_46_v) + t_5)) ? t_2 : ((t_2 != t_2) ? Float32(Float32(Float32((floor(h) ^ Float32(2.0)) * dX_46_v) * dX_46_v) + t_5) : max(Float32(Float32(Float32((floor(h) ^ Float32(2.0)) * dX_46_v) * dX_46_v) + t_5), t_2))))) * t_4); else tmp = Float32(Float32(Float32(1.0) / sqrt(((Float32(Float32(t_3 * t_3) + t_5) != Float32(Float32(t_3 * t_3) + t_5)) ? t_2 : ((t_2 != t_2) ? Float32(Float32(t_3 * t_3) + t_5) : max(Float32(Float32(t_3 * t_3) + t_5), t_2))))) * t_1); end return tmp end
function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = dY_46_v * floor(h); t_1 = dY_46_u * floor(w); t_2 = (t_0 * t_0) + (t_1 * t_1); t_3 = floor(h) * dX_46_v; t_4 = floor(w) * dX_46_u; t_5 = t_4 * t_4; tmp = single(0.0); if (((t_3 ^ single(2.0)) + (t_4 ^ single(2.0))) >= (t_1 ^ single(2.0))) tmp = (single(1.0) / sqrt(max(((((floor(h) ^ single(2.0)) * dX_46_v) * dX_46_v) + t_5), t_2))) * t_4; else tmp = (single(1.0) / sqrt(max(((t_3 * t_3) + t_5), t_2))) * t_1; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_1 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_2 := t\_0 \cdot t\_0 + t\_1 \cdot t\_1\\
t_3 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_4 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_5 := t\_4 \cdot t\_4\\
\mathbf{if}\;{t\_3}^{2} + {t\_4}^{2} \geq {t\_1}^{2}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(\left({\left(\left\lfloor h\right\rfloor \right)}^{2} \cdot dX.v\right) \cdot dX.v + t\_5, t\_2\right)}} \cdot t\_4\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_3 \cdot t\_3 + t\_5, t\_2\right)}} \cdot t\_1\\
\end{array}
\end{array}
Initial program 77.1%
Taylor expanded in dY.u around inf
*-commutativeN/A
unpow2N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3264.7
Applied rewrites64.7%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3264.7
Applied rewrites64.7%
lift-+.f32N/A
+-commutativeN/A
lift-*.f32N/A
pow2N/A
lower-pow.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-+.f3264.7
Applied rewrites64.7%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
swap-sqrN/A
unpow2N/A
lift-pow.f32N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f3264.8
Applied rewrites64.8%
Final simplification64.8%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(t_1 (* dY.v (floor h)))
(t_2 (* dY.u (floor w)))
(t_3 (* t_2 t_2))
(t_4 (* (floor h) dX.v))
(t_5 (+ (* t_4 t_4) (* t_0 t_0))))
(if (>= (+ (pow t_4 2.0) (pow t_0 2.0)) (pow t_2 2.0))
(* (/ 1.0 (sqrt (fmax t_5 (+ (* t_1 t_1) t_3)))) t_0)
(*
(/ 1.0 (sqrt (fmax t_5 (+ (* (* dY.v dY.v) (pow (floor h) 2.0)) t_3))))
t_2))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(w) * dX_46_u;
float t_1 = dY_46_v * floorf(h);
float t_2 = dY_46_u * floorf(w);
float t_3 = t_2 * t_2;
float t_4 = floorf(h) * dX_46_v;
float t_5 = (t_4 * t_4) + (t_0 * t_0);
float tmp;
if ((powf(t_4, 2.0f) + powf(t_0, 2.0f)) >= powf(t_2, 2.0f)) {
tmp = (1.0f / sqrtf(fmaxf(t_5, ((t_1 * t_1) + t_3)))) * t_0;
} else {
tmp = (1.0f / sqrtf(fmaxf(t_5, (((dY_46_v * dY_46_v) * powf(floorf(h), 2.0f)) + t_3)))) * t_2;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(w) * dX_46_u) t_1 = Float32(dY_46_v * floor(h)) t_2 = Float32(dY_46_u * floor(w)) t_3 = Float32(t_2 * t_2) t_4 = Float32(floor(h) * dX_46_v) t_5 = Float32(Float32(t_4 * t_4) + Float32(t_0 * t_0)) tmp = Float32(0.0) if (Float32((t_4 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) >= (t_2 ^ Float32(2.0))) tmp = Float32(Float32(Float32(1.0) / sqrt(((t_5 != t_5) ? Float32(Float32(t_1 * t_1) + t_3) : ((Float32(Float32(t_1 * t_1) + t_3) != Float32(Float32(t_1 * t_1) + t_3)) ? t_5 : max(t_5, Float32(Float32(t_1 * t_1) + t_3)))))) * t_0); else tmp = Float32(Float32(Float32(1.0) / sqrt(((t_5 != t_5) ? Float32(Float32(Float32(dY_46_v * dY_46_v) * (floor(h) ^ Float32(2.0))) + t_3) : ((Float32(Float32(Float32(dY_46_v * dY_46_v) * (floor(h) ^ Float32(2.0))) + t_3) != Float32(Float32(Float32(dY_46_v * dY_46_v) * (floor(h) ^ Float32(2.0))) + t_3)) ? t_5 : max(t_5, Float32(Float32(Float32(dY_46_v * dY_46_v) * (floor(h) ^ Float32(2.0))) + t_3)))))) * t_2); end return tmp end
function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(w) * dX_46_u; t_1 = dY_46_v * floor(h); t_2 = dY_46_u * floor(w); t_3 = t_2 * t_2; t_4 = floor(h) * dX_46_v; t_5 = (t_4 * t_4) + (t_0 * t_0); tmp = single(0.0); if (((t_4 ^ single(2.0)) + (t_0 ^ single(2.0))) >= (t_2 ^ single(2.0))) tmp = (single(1.0) / sqrt(max(t_5, ((t_1 * t_1) + t_3)))) * t_0; else tmp = (single(1.0) / sqrt(max(t_5, (((dY_46_v * dY_46_v) * (floor(h) ^ single(2.0))) + t_3)))) * t_2; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_2 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_3 := t\_2 \cdot t\_2\\
t_4 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_5 := t\_4 \cdot t\_4 + t\_0 \cdot t\_0\\
\mathbf{if}\;{t\_4}^{2} + {t\_0}^{2} \geq {t\_2}^{2}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_5, t\_1 \cdot t\_1 + t\_3\right)}} \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_5, \left(dY.v \cdot dY.v\right) \cdot {\left(\left\lfloor h\right\rfloor \right)}^{2} + t\_3\right)}} \cdot t\_2\\
\end{array}
\end{array}
Initial program 77.1%
Taylor expanded in dY.u around inf
*-commutativeN/A
unpow2N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3264.7
Applied rewrites64.7%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3264.7
Applied rewrites64.7%
lift-+.f32N/A
+-commutativeN/A
lift-*.f32N/A
pow2N/A
lower-pow.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-+.f3264.7
Applied rewrites64.7%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
unpow-prod-downN/A
lift-pow.f32N/A
lower-*.f32N/A
pow2N/A
lower-*.f3264.7
Applied rewrites64.7%
Final simplification64.7%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(t_1 (* dY.u (floor w)))
(t_2 (* t_1 t_1))
(t_3 (* (floor h) dX.v))
(t_4 (* dY.v (floor h)))
(t_5 (+ (* t_3 t_3) (* t_0 t_0))))
(if (>= (+ (pow t_3 2.0) (pow t_0 2.0)) (pow t_1 2.0))
(*
(/
1.0
(sqrt (fmax t_5 (+ (* (* (* dY.v dY.v) (floor h)) (floor h)) t_2))))
t_0)
(* (/ 1.0 (sqrt (fmax t_5 (+ (* t_4 t_4) t_2)))) t_1))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(w) * dX_46_u;
float t_1 = dY_46_u * floorf(w);
float t_2 = t_1 * t_1;
float t_3 = floorf(h) * dX_46_v;
float t_4 = dY_46_v * floorf(h);
float t_5 = (t_3 * t_3) + (t_0 * t_0);
float tmp;
if ((powf(t_3, 2.0f) + powf(t_0, 2.0f)) >= powf(t_1, 2.0f)) {
tmp = (1.0f / sqrtf(fmaxf(t_5, ((((dY_46_v * dY_46_v) * floorf(h)) * floorf(h)) + t_2)))) * t_0;
} else {
tmp = (1.0f / sqrtf(fmaxf(t_5, ((t_4 * t_4) + t_2)))) * t_1;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(w) * dX_46_u) t_1 = Float32(dY_46_u * floor(w)) t_2 = Float32(t_1 * t_1) t_3 = Float32(floor(h) * dX_46_v) t_4 = Float32(dY_46_v * floor(h)) t_5 = Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0)) tmp = Float32(0.0) if (Float32((t_3 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) >= (t_1 ^ Float32(2.0))) tmp = Float32(Float32(Float32(1.0) / sqrt(((t_5 != t_5) ? Float32(Float32(Float32(Float32(dY_46_v * dY_46_v) * floor(h)) * floor(h)) + t_2) : ((Float32(Float32(Float32(Float32(dY_46_v * dY_46_v) * floor(h)) * floor(h)) + t_2) != Float32(Float32(Float32(Float32(dY_46_v * dY_46_v) * floor(h)) * floor(h)) + t_2)) ? t_5 : max(t_5, Float32(Float32(Float32(Float32(dY_46_v * dY_46_v) * floor(h)) * floor(h)) + t_2)))))) * t_0); else tmp = Float32(Float32(Float32(1.0) / sqrt(((t_5 != t_5) ? Float32(Float32(t_4 * t_4) + t_2) : ((Float32(Float32(t_4 * t_4) + t_2) != Float32(Float32(t_4 * t_4) + t_2)) ? t_5 : max(t_5, Float32(Float32(t_4 * t_4) + t_2)))))) * t_1); end return tmp end
function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(w) * dX_46_u; t_1 = dY_46_u * floor(w); t_2 = t_1 * t_1; t_3 = floor(h) * dX_46_v; t_4 = dY_46_v * floor(h); t_5 = (t_3 * t_3) + (t_0 * t_0); tmp = single(0.0); if (((t_3 ^ single(2.0)) + (t_0 ^ single(2.0))) >= (t_1 ^ single(2.0))) tmp = (single(1.0) / sqrt(max(t_5, ((((dY_46_v * dY_46_v) * floor(h)) * floor(h)) + t_2)))) * t_0; else tmp = (single(1.0) / sqrt(max(t_5, ((t_4 * t_4) + t_2)))) * t_1; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_2 := t\_1 \cdot t\_1\\
t_3 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_4 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_5 := t\_3 \cdot t\_3 + t\_0 \cdot t\_0\\
\mathbf{if}\;{t\_3}^{2} + {t\_0}^{2} \geq {t\_1}^{2}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_5, \left(\left(dY.v \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right) \cdot \left\lfloor h\right\rfloor + t\_2\right)}} \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_5, t\_4 \cdot t\_4 + t\_2\right)}} \cdot t\_1\\
\end{array}
\end{array}
Initial program 77.1%
Taylor expanded in dY.u around inf
*-commutativeN/A
unpow2N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3264.7
Applied rewrites64.7%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3264.7
Applied rewrites64.7%
lift-+.f32N/A
+-commutativeN/A
lift-*.f32N/A
pow2N/A
lower-pow.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-+.f3264.7
Applied rewrites64.7%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
unpow-prod-downN/A
unpow2N/A
associate-*l*N/A
lower-*.f32N/A
lower-*.f32N/A
pow2N/A
lower-*.f3264.7
Applied rewrites64.7%
Final simplification64.7%
herbie shell --seed 2024308
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:name "Anisotropic x16 LOD (line direction, u)"
:precision binary32
:pre (and (and (and (and (and (and (and (<= 1.0 w) (<= w 16384.0)) (and (<= 1.0 h) (<= h 16384.0))) (and (<= 1e-20 (fabs dX.u)) (<= (fabs dX.u) 1e+20))) (and (<= 1e-20 (fabs dX.v)) (<= (fabs dX.v) 1e+20))) (and (<= 1e-20 (fabs dY.u)) (<= (fabs dY.u) 1e+20))) (and (<= 1e-20 (fabs dY.v)) (<= (fabs dY.v) 1e+20))) (== maxAniso 16.0))
(if (>= (+ (* (* (floor w) dX.u) (* (floor w) dX.u)) (* (* (floor h) dX.v) (* (floor h) dX.v))) (+ (* (* (floor w) dY.u) (* (floor w) dY.u)) (* (* (floor h) dY.v) (* (floor h) dY.v)))) (* (/ 1.0 (sqrt (fmax (+ (* (* (floor w) dX.u) (* (floor w) dX.u)) (* (* (floor h) dX.v) (* (floor h) dX.v))) (+ (* (* (floor w) dY.u) (* (floor w) dY.u)) (* (* (floor h) dY.v) (* (floor h) dY.v)))))) (* (floor w) dX.u)) (* (/ 1.0 (sqrt (fmax (+ (* (* (floor w) dX.u) (* (floor w) dX.u)) (* (* (floor h) dX.v) (* (floor h) dX.v))) (+ (* (* (floor w) dY.u) (* (floor w) dY.u)) (* (* (floor h) dY.v) (* (floor h) dY.v)))))) (* (floor w) dY.u))))